To remove information on a part which is not concerned in the improvement of recognizing performance.

This speech recognizing method which has a feature quantity converting means for converting input data into a feature quantity for discriminating a phonemic class discriminated in a module and a means for discriminating the phonemic class in the module by using the converted feature quantity and is characterized by that the input data are converted into the feature quantity by removing information which does not represents the features of the phonemic class discriminated in the module improves the recognition precision by allowing the feature quantity converting means to use as a feature quantity only the information where the feature of the phonemic class appears by removing the information which does not represent the feature of the phonemic class discriminated in the module. Further, the number of phonemic classes discriminated in respective modules is decreased by constitution and then information representing the features of the phonemic classes is narrowed down, so the recognition precision is improved.